Function definitions, setup
library(stringr)
library(ggplot2)
library(dplyr)
library(magrittr)
library(corrplot)
library(robust)
library(ggpubr)
library(fit.models)
library(matrixStats)
library(scatterplot3d)
Global setting
# Use these for more than 10 algorithms
colors_20 = c(
"#1f77b4", "#aec7e8", "#ff7f0e", "#ffbb78", "#2ca02c", "#98df8a", "#d62728", "#ff9896",
"#9467bd", "#c5b0d5", "#8c564b", "#c49c94", "#e377c2", "#f7b6d2", "#7f7f7f", "#c7c7c7",
"#bcbd22", "#dbdb8d", "#17becf", "#9edae5")
# Use these for 10 or fewer algorithms
colors_10 = c("#1f77b4", "#ff7f0e", "#2ca02c", "#d62728", "#9467bd", "#8c564b", "#e377c2", "#7f7f7f", "#bcbd22", "#17becf")
# symbol = c(21:25, 1, 0, 5, 2, 6,
# 7:14, 3, 4)
symbol = c(15:18, 3, 4, 5, 7:14)
full_baseline = c("LR", "DecisionTree",
# "Calders",
# "Kamishima",
"GP", "SVM", "DT",
"ZafarEqOpp")
# "Feldman-SVM", "Feldman-GaussianNB", "Feldman-LR", "Feldman-DecisionTree",
catscale10 = scale_colour_manual(values=colors_10)
catscale10_2 = scale_fill_manual(values=colors_10)
catscale20 = scale_colour_manual(values=colors_20)
catscale20_2 = scale_fill_manual(values=colors_20)
true_name = character()
false_name = character()
true_tnr0 = character()
false_tnr0 = character()
true_tnr1 = character()
false_tnr1 = character()
true_tpr = character()
false_tpr = character()
true_tnr = character()
false_tnr = character()
for (i in 1: 4) {
for (k in 1:4) {
true_name = c(true_name, paste("FairGPopp*, 0-TNR=", i*0.1 + 0.5, ", 1-TNR=",
i*0.1 + 0.5, ", TPR=",
k*0.1 + 0.5, sep = ""))
true_tnr0 = c(true_tnr0, paste("FairGPopp*, 0-TNR=", i*0.1 + 0.5, sep = ""))
true_tnr1 = c(true_tnr1, paste("FairGPopp*, 1-TNR=", i*0.1 + 0.5, sep = ""))
true_tnr = c(true_tnr, paste("FairGPopp*, TNR=", i*0.1 + 0.5, sep = ""))
true_tpr = c(true_tpr, paste("FairGPopp*, TPR=", k*0.1 + 0.5, sep = ""))
false_name = c(false_name, paste("FairGPopp, 0-TNR=", i*0.1 + 0.5, ", 1-TNR=",
i*0.1 + 0.5, ", TPR=",
k*0.1 + 0.5, sep = ""))
false_tnr0 = c(false_tnr0, paste("FairGPopp, 0-TNR=", i*0.1 + 0.5, sep = ""))
false_tnr1 = c(false_tnr1, paste("FairGPopp, 1-TNR=", i*0.1 + 0.5, sep = ""))
false_tnr = c(false_tnr, paste("FairGPopp, TNR=", i*0.1 + 0.5, sep = ""))
false_tpr = c(false_tpr, paste("FairGPopp, TPR=", k*0.1 + 0.5, sep = ""))
}
}
algos_used_full = c(false_name)
tnr0_full = c(false_tnr0)
tnr1_full = c(false_tnr1)
tnr_full = c(false_tnr)
tpr_full = c(false_tpr)
Function
make_scatter_figure = function(name, var1="CV", var2="accuracy", algos=algos_used_full, view="algorithm",
setTNR0=tnr0_full, setTNR1=tnr1_full, setTPRset=tpr_full, tnr_show=tnr_full,
display="off", statistic = "off", baseline=full_baseline) {
x_var = as.name(var1)
y_var = as.name(var2)
target_view = as.name(view)
df = read.csv(str_c(name, "_numerical-binsensitive.csv"), check.names=FALSE)%>%
filter(algorithm %in% algos, TNR0set%in% setTNR0, TNR1set%in% setTNR1, TPRset %in% setTPRset)
df = df[df$TNR0num == df$TNR1num, ]
# df[["race-TPRDiff"]] = 1 - df[["race-TPRDiff"]]
# df[["sex-TPRDiff"]] = 1 - df[["sex-TPRDiff"]]
# df[["race-TNRDiff"]] = 1 - df[["race-TNRDiff"]]
# df[["sex-TNRDiff"]] = 1 - df[["sex-TNRDiff"]]
df[["race-AAD"]] = (abs(1 - df[["race-TPRDiff"]]) + abs(1 - df[["race-TNRDiff"]]))/2
df[["sex-AAD"]] = ( abs(1 -df[["sex-TPRDiff"]]) + abs(1 - df[["sex-TNRDiff"]]))/2
switch (statistic,
median = {
switch (view,
algorithm ={
num = length(algos)
variable = algos
tnr_show_view = algos
},
TPRset = {
num = length(unique(setTPRset))
variable = unique(setTPRset)
tnr_show_view = variable
},
TNR0set = {
num = length(unique(setTNR0))
variable = unique(setTNR0)
tnr_show_view = unique(tnr_show)
},
TNR1set = {
num = length(unique(setTNR1))
variable = unique(setTNR1)
tnr_show_view = unique(tnr_show)
}
)
new_df = matrix(nrow=num,ncol=2, dimnames =list(1:num, c(var1, var2)))
for(j in 1:num){
temp_data = df[which(df[[view]]==variable[j]),c(var1, var2)]
new_df[j, ] = colMedians(data.matrix(temp_data))
}
dataFrame <- data.frame(new_df)
names(dataFrame) <- c(var1,var2)
dataFrame[[view]] <- unique(tnr_show)
},
mean = {
switch (view,
algorithm ={
num = length(algos)
variable = algos
tnr_show_view = algos
},
TPRset = {
num = length(unique(setTPRset))
variable = unique(setTPRset)
tnr_show_view = variable
},
TNR0set = {
num = length(unique(setTNR0))
variable = unique(setTNR0)
tnr_show_view = unique(tnr_show)
},
TNR1set = {
num = length(unique(setTNR1))
variable = unique(setTNR1)
tnr_show_view = unique(tnr_show)
}
)
new_df = matrix(nrow=num,ncol=2, dimnames =list(1:num, c(var1, var2)))
for(j in 1:num){
temp_data = df[which(df[[view]]==variable[j]),c(var1, var2)]
new_df[j, ] = colMeans(data.matrix(temp_data))
}
dataFrame <- data.frame(new_df)
names(dataFrame) <- c(var1,var2)
dataFrame[[view]] <- tnr_show_view
},
mean_of_repeats = {
num = length(algos)
switch (view,
algorithm ={
variable = algos
tnr_show_view = algos
},
TPRset = {
variable = setTPRset
tnr_show_view = variable
},
TNR0set = {
variable = setTNR0
tnr_show_view = tnr_show
},
TNR1set = {
variable = setTNR1
tnr_show_view = tnr_show
}
)
new_df = matrix(nrow=num,ncol=2, dimnames =list(1:num, c(var1, var2)))
for(j in 1:num){
temp_data = df[which(df$algorithm==algos[j]),c(var1, var2)]
new_df[j, ] = colMeans(data.matrix(temp_data))
}
dataFrame <- data.frame(new_df)
names(dataFrame) <- c(var1,var2)
#dataFrame$algorithm <- algos
dataFrame[[view]] = tnr_show_view
},
off = {
dataFrame = df
}
)
color_used = catscale10
color_used_2 = catscale10_2
if(is.character(baseline)){
df_baseline = read.csv(str_c(name, "_baseline.csv"), check.names=FALSE)%>%
filter(algorithm %in% baseline)
df_baseline [["race-AAD"]] = (abs(1 - df_baseline [["race-TPRDiff"]]) + abs(1 - df_baseline [["race-TNRDiff"]]))/2
df_baseline [["sex-AAD"]] = (abs(1 - df_baseline [["sex-TPRDiff"]]) + abs(1 - df_baseline [["sex-TNRDiff"]]))/2
num_baseline = length(baseline)
new_df_baseline = matrix(nrow=num_baseline,ncol=2, dimnames =list(1:num_baseline, c(var1, var2)))
for(j in 1:num_baseline){
temp_baseline = df_baseline[which(df_baseline$algorithm==baseline[j]),c(var1, var2)]
new_df_baseline[j, ] = colMeans(data.matrix(temp_baseline))
}
dataFrame_baseline <- data.frame(new_df_baseline)
names(dataFrame_baseline) <- c(var1,var2)
#dataFrame$algorithm <- algos
dataFrame_baseline[[view]] = baseline
color_used = catscale20
color_used_2 = catscale20_2
}
dataFrame = rbind(dataFrame, dataFrame_baseline)
dataFrame[[view]] = factor(dataFrame[[view]], levels= unique(dataFrame[[view]]))
# Scatterplot
fig_title = paste(name, "(", statistic, ")", sep="")
switch(display,
off = ggplot(dataFrame, aes_q(x=x_var, y=y_var, colour=target_view, bg=target_view)) + color_used + color_used_2 +
geom_point(size=3, aes_q(shape=target_view)) + scale_shape_manual(values=symbol) +
labs(y=var2, x=var1),
title = ggplot(dataFrame, aes_q(x=x_var, y=y_var, colour=target_view, bg=target_view)) + color_used + color_used_2 +
geom_point(size=3, aes_q(shape=target_view)) + scale_shape_manual(values=symbol) +
labs(y=var2, x=var1, title=fig_title),
captain = ggplot(dataFrame, aes_q(x=x_var, y=y_var, colour=target_view, bg=target_view)) + color_used + color_used_2 +
geom_point(size=3, aes_q(shape=target_view)) + scale_shape_manual(values=symbol) +
labs(y=var2, x=var1, caption = str_c(name, "_numerical-binsensitive.csv")),
all = ggplot(dataFrame, aes_q(x=x_var, y=y_var, colour=target_view, bg=target_view)) + color_used + color_used_2 +
geom_point(size=3, aes_q(shape=target_view)) + scale_shape_manual(values=symbol) +
labs(y=var2, x=var1, title=fig_title)
)
}
Users setting
file_name = list("propublica-recidivism_race","propublica-recidivism_sex")
file_attribute = list( "race", "sex")
N_file = length(file_name)
Algorithm accuracy vs fairness (group-TPRDiff) all points
variable_1 = "TPRDiff"
add_for_1 = 1
variable_2 = "TNRDiff"
add_for_2 =1
variable_3 = "AAD"
add_for_3 =1
target = "accuracy"
add_for_target = 0
view_set = c("TPRset","TNR0set")
for (i in 1:N_file) {
for (view_from in view_set){
if(add_for_target){
target_name = paste(file_attribute[i], "-", target, sep="")
} else {
target_name = target
}
if(add_for_1){
var1_name = paste(file_attribute[i], "-", variable_1, sep="")
} else {
var1_name = variable_1
}
q1 = make_scatter_figure(file_name[i], var1=var1_name, var2=target_name,
view=view_from, display = "all", statistic = "mean_of_repeats")
if(add_for_2){
var2_name = paste(file_attribute[i], "-", variable_2, sep="")
} else {
var2_name = variable_2
}
q2 = make_scatter_figure(file_name[i], var1=var2_name, var2=target_name,
view=view_from, display = "off", statistic = "mean_of_repeats")
q = ggarrange(q1, q2, ncol=1, nrow=2, common.legend = TRUE, legend="right")
print(q)
if(add_for_3){
var3_name = paste(file_attribute[i], "-", variable_3, sep="")
} else {
var3_name = variable_3
}
q3 = make_scatter_figure(file_name[i], var1=var3_name, var2=target_name,
view=view_from, display = "all", statistic = "mean_of_repeats")
print(q3)
# export_name = paste(file_name[i], "_opp.eps", sep="")
# ggsave(export_name, q)
}
}
Algorithm accuracy vs fairness (group-TPRDiff) in terms of mean
for (i in 1:N_file) {
for (view_from in view_set){
q1 = make_scatter_figure(file_name[i], var1=var1_name, var2=target_name,
view=view_from, display = "all", statistic = "mean")
q2 = make_scatter_figure(file_name[i], var1=var2_name, var2=target_name,
view=view_from, display = "off", statistic = "mean")
q = ggarrange(q1, q2, ncol=1, nrow=2, common.legend = TRUE, legend="right")
print(q)
q3 = make_scatter_figure(file_name[i], var1=var3_name, var2=target_name,
view=view_from, display = "all", statistic = "mean")
print(q3)
export_name = paste(file_name[i], "_odds.eps", sep="")
# ggsave(export_name, q)
}
}
